"Convolutional Networks: Unleashing the Potential of Machine Learning for Robust Perception...
-
Upload
embedded-vision-alliance -
Category
Technology
-
view
47 -
download
1
Transcript of "Convolutional Networks: Unleashing the Potential of Machine Learning for Robust Perception...
Y LeCun
Convolutional Networks:Unleashing the Potential of
Machine Learning forRobust Perception Systems
Yann LeCunFacebook AI Research &Center for Data Science, [email protected]://yann.lecun.com
Y LeCun55 years of hand-crafted features
The traditional model of pattern recognition (since the late 50's)Fixed/engineered features (or fixed kernel) + trainable classifier
Perceptron
“Simple” Trainable Classifier
hand-craftedFeature Extractor
Y LeCunArchitecture of “Mainstream” Pattern Recognition Systems
Modern architecture for pattern recognitionSpeech recognition: early 90's – 2011
Object Recognition: 2006 - 2012
fixed unsupervised supervised
ClassifierMFCC Mix of Gaussians
ClassifierSIFTHoG
K-meansSparse Coding
Pooling
fixed unsupervised supervised
Low-levelFeatures
Mid-levelFeatures
Y LeCunDeep Learning = Learning Hierarchical Representations
Traditional Pattern Recognition: Fixed/Handcrafted Feature Extractor
Trainable Classifier
Feature Extractor
Mainstream Modern Pattern Recognition: Unsupervised mid-level features
Trainable Classifier
Feature Extractor
Mid-LevelFeatures
Deep Learning: Representations are hierarchical and trained
Trainable Classifier
Low-LevelFeatures
Mid-LevelFeatures
High-LevelFeatures
Y LeCunDeep Learning = Learning Hierarchical Representations
It's deep if it has more than one stage of non-linear feature transformation
Trainable Classifier
Low-LevelFeature
Mid-LevelFeature
High-LevelFeature
Feature visualization of convolutional net trained on ImageNet from [Zeiler & Fergus 2013]
Y LeCunTrainable Feature Hierarchy
Hierarchy of representations with increasing level of abstraction
Each stage is a kind of trainable feature transform
Image recognitionPixel edge texton motif part object→ → → → →
TextCharacter word word group clause sentence story→ → → → →
SpeechSample spectral band sound … phone phoneme word→ → → → → →
Y LeCun
Learning Representations: a challenge forML, CV, AI, Neuroscience, Cognitive Science...
How do we learn representations of the perceptual world?How can a perceptual system build itself by looking at the world?How much prior structure is necessary
ML/AI: how do we learn features or feature hierarchies?What is the fundamental principle? What is the learning algorithm? What is the architecture?
Neuroscience: how does the cortex learn perception?Does the cortex “run” a single, general learning algorithm? (or a small number of them)
CogSci: how does the mind learn abstract concepts on top of less abstract ones?
Deep Learning addresses the problem of learning hierarchical representations with a single algorithm
or perhaps with a few algorithms
Trainable FeatureTransform
Trainable FeatureTransform
Trainable FeatureTransform
Trainable FeatureTransform
Y LeCunThe Mammalian Visual Cortex is Hierarchical
[picture from Simon Thorpe]
[Gallant & Van Essen]
The ventral (recognition) pathway in the visual cortex has multiple stagesRetina - LGN - V1 - V2 - V4 - PIT - AIT ....Lots of intermediate representations
Y LeCunLet's be inspired by nature, but not too much
It's nice imitate Nature,But we also need to understand
How do we know which details are important?
Which details are merely the result of evolution, and the constraints of biochemistry?
For airplanes, we developed aerodynamics and compressible fluid dynamics.
We figured that feathers and wing flapping weren't crucial
QUESTION: What is the equivalent of aerodynamics for understanding intelligence?
L'Avion III de Clément Ader, 1897(Musée du CNAM, Paris)
His “Eole” took off from the ground in 1890,
13 years before the Wright Brothers, but you
probably never heard of it (unless you are french).
Y LeCun
Shallow vs Deep == lookup table vs multi-step algorithm
“shallow & wide” vs “deep and narrow” == “more memory” vs “more time”Look-up table vs algorithmFew functions can be computed in two steps without an exponentially large lookup tableUsing more than 2 steps can reduce the “memory” by an exponential factor.
Step 1
Step 2
Step 3
Step 4
Step 1 (look up table/templates)
Step 2
Y LeCun
What AreGood Feature?
Y LeCun
Discovering the Hidden Structure in High-Dimensional DataThe manifold hypothesis
Learning Representations of Data:
Discovering & disentangling the independent explanatory factors
The Manifold Hypothesis:Natural data lives in a low-dimensional (non-linear) manifold
Because variables in natural data are mutually dependent
Y LeCunDiscovering the Hidden Structure in High-Dimensional Data
Example: all face images of a person1000x1000 pixels = 1,000,000 dimensions
But the face has 3 cartesian coordinates and 3 Euler angles
And humans have less than about 50 muscles in the face
Hence the manifold of face images for a person has <56 dimensions
The perfect representations of a face image:Its coordinates on the face manifold
Its coordinates away from the manifold
We do not have good and general methods to learn functions that turns an image into this kind of representation
IdealFeature
Extractor [1 . 2−30 . 2
−2 .. .]
Face/not facePoseLightingExpression-----
Y LeCunBasic Idea for Invariant Feature Learning
Embed the input non-linearly into a high(er) dimensional spaceIn the new space, things that were non separable may become separable
Pool regions of the new space togetherBringing together things that are semantically similar. Like pooling.
Non-LinearFunction
PoolingOr
Aggregation
Inputhigh-dim
Unstable/non-smooth features
Stable/invariantfeatures
Y LeCunSparse Non-Linear Expansion → Pooling
Use clustering to break things apart, pool together similar things
Clustering,Quantization,Sparse Coding
Pooling.Aggregation
Y LeCun
Overall Architecture: multiple stages of Normalization → Filter Bank → Non-Linearity → Pooling
Normalization: variation on whitening (optional)
– Subtractive: average removal, high pass filtering– Divisive: local contrast normalization, variance normalizationFilter Bank: dimension expansion, projection on overcomplete basisNon-Linearity: sparsification, saturation, lateral inhibition....
– Rectification (ReLU), Component-wise shrinkage, tanh,..
Pooling: aggregation over space or feature type
– Max, Lp norm, log prob.
MAX :Max i(X i) ; L p :p√ X ip ; PROB :
1b
log(∑i ebX i)
Classifierfeature
Pooling
Non-
Linear
Filter
Bank Norm
feature
Pooling
Non-
Linear
Filter
Bank Norm
ReLU (x )=max (x ,0)
Y LeCun
ConvolutionalNetworks
Y LeCunConvolutional Network
[LeCun et al. NIPS 1989]
Filter Bank +non-linearity
Filter Bank +non-linearity
Pooling
Pooling
Filter Bank +non-linearity
Y LeCunEarly Hierarchical Feature Models for Vision
[Hubel & Wiesel 1962]: simple cells detect local features
complex cells “pool” the outputs of simple cells within a retinotopic neighborhood.
Cognitron & Neocognitron [Fukushima 1974-1982]
pooling subsampling
“Simple cells”“Complex cells”
Multiple convolutions
Y LeCun
The Convolutional Net Model (Multistage Hubel-Wiesel system)
pooling subsampling
“Simple cells”“Complex cells”
Multiple convolutions
Retinotopic Feature Maps
[LeCun et al. 89][LeCun et al. 98]
Training is supervisedWith stochastic gradient descent
Y LeCunSupervised Training: Back Propgation (with tricks)
Y LeCunConvolutional Network (vintage 1990)
filters → tanh → average-tanh → filters → tanh → average-tanh → filters → tanh
Curvedmanifold
Flattermanifold
Y LeCun
“Mainstream” object recognition pipeline 2006-2012: somewhat similar to ConvNets
Fixed Features + unsupervised mid-level features + simple classifier SIFT + Vector Quantization + Pyramid pooling + SVM
[Lazebnik et al. CVPR 2006]
SIFT + Local Sparse Coding Macrofeatures + Pyramid pooling + SVM
[Boureau et al. ICCV 2011]
SIFT + Fisher Vectors + Deformable Parts Pooling + SVM
[Perronin et al. 2012]
Oriented
Edges
Winner
Takes AllHistogram
(sum)
Filter
Bank
feature
Pooling
Non-
Linearity
Filter
Bank
feature
Pooling
Non-
LinearityClassifier
Fixed (SIFT/HoG/...)
K-means
Sparse CodingSpatial Max
Or averageAny simple
classifier
Unsupervised Supervised
Y LeCun
Convolutional NetworksIn
Visual Object Recognition
Y LeCunObject Recognition [Krizhevsky, Sutskever, Hinton 2012]
CONV 11x11/ReLU 96fmLOCAL CONTRAST NORM
MAX POOL 2x2sub
FULL 4096/ReLUFULL CONNECT
CONV 11x11/ReLU 256fmLOCAL CONTRAST NORM
MAX POOLING 2x2sub
CONV 3x3/ReLU 384fmCONV 3x3ReLU 384fm
CONV 3x3/ReLU 256fmMAX POOLING
FULL 4096/ReLU
Won the 2012 ImageNet LSVRC. 60 Million parameters, 832M MAC ops4M
16M37M
442K
1.3M884K
307K
35K
4Mflop
16M37M
74M
224M149M
223M
105M
Y LeCunObject Recognition [Krizhevsky, Sutskever, Hinton 2012]
Method: large convolutional net650K neurons, 832M synapses, 60M parameters
Trained with backprop on GPU
Trained “with all the tricks Yann came up with in the last 20 years, plus dropout” (Hinton, NIPS 2012)
Rectification, contrast normalization,...
Error rate: 15% (whenever correct class isn't in top 5)Previous state of the art: 25% error
A REVOLUTION IN COMPUTER VISION
Acquired by Google in Jan 2013Deployed in Google+ Photo Tagging in May 2013
Y LeCunConvNet-Based Google+ Photo Tagger
Searched my personal collection for “bird”
SamyBengio???
Y LeCunNYU ConvNet Trained on ImageNet: OverFeat
[Sermanet et al. arXiv:1312.6229]
Trained on GPU using Torch7
Uses a number of new tricks
Classification 1000 categories:13.8% error (top 5) with an ensemble of 7 networks (Krizhevsky: 15%)15.4% error (top 5) with a single network (Krizhevksy: 18.2%)
Classification+Localization30% error (Krizhevsky: 34%)
Detection (200 categories)19% correct
Dowloadable code (running, no training)Search for “overfeat NYU” on Googlehttp://cilvr.nyu.edu software→
CONV 7x7/ReLU 96fm
MAX POOL 3x3sub
FULL 4096/ReLUFULL 1000/Softmax
CONV 7x7/ReLU 256fmMAX POOLING 2x2sub
CONV 3x3/ReLU 384fmCONV 3x3ReLU 384fm
CONV 3x3/ReLU 256fmMAX POOLING 3x3sub
FULL 4096/ReLU
Y LeCunKernels: Layer 1 (7x7)and Layer 2 (7x7)
Layer 1: 3x96 kernels, RGB->96 feature maps, 7x7 Kernels, stride 2
Layer 2: 96x256 kernels, 7x7
Y LeCunKernels: Layer 1 (11x11)
Layer 1: 3x96 kernels, RGB->96 feature maps, 11x11 Kernels, stride 4
Y LeCunImageNet 2013: Classification
Give the name of the dominant object in the image
Top-5 error rates: if correct class is not in top 5, count as error
(NYU Teams in Purple)
Y LeCunClassification+Localization. Results
Y LeCunClassification+Localization. Error Rates
It's best to propose several categories for the same windowOne of them might be right
Y LeCun
Classification + Localization: multiscale sliding window
Apply convnet with a sliding window over the image at multiple scales
Important note: it's very cheap to slide a convnet over an imageJust compute the convolutions over the whole image and replicate the fully-connected layers
Y LeCun
96x96
input:120x120
output: 3x3
Traditional Detectors/Classifiers must be applied to every location on a large input image, at multiple scales. Convolutional nets can replicated over large images very cheaply. Simply apply the convolutions to the entire image and spatially replicate the fully-connected layers
Applying a ConvNet on Sliding Windows is Very Cheap!
Y LeCun
Classification + Localization: sliding window + bounding box regression
Apply convnet with a sliding window over the image at multiple scales
For each window, predict a class and bounding box parametersEvenif the object is not completely contained in the viewing window, the convnet can predict where it thinks the object is.
Y LeCun
Classification + Localization: sliding window + bounding box regression + bbox voting
Apply convnet with a sliding window over the image at multiple scales
For each window, predict a class and bounding box parameters
Compute an “average” bounding box, weighted by scores
Y LeCunLocalization: Sliding Window + bbox vote + multiscale
Y LeCunDetection / Localization
OverFeat • Pierre Sermanet • New York University
Y LeCunDetection / Localization
OverFeat • Pierre Sermanet • New York University
Y LeCunDetection / Localization
OverFeat • Pierre Sermanet • New York University
Y LeCunDetection / Localization
OverFeat • Pierre Sermanet • New York University
Y LeCunDetection / Localization
OverFeat • Pierre Sermanet • New York University
Y LeCunDetection: Examples
200 broad categories
There is a penalty for false positives
Some examples are easy some are impossible/ambiguous
Some classes are well detected
Burritos?
Y LeCunDetection: Examples
Groundtruth is sometimes ambiguous or incomplete
Large overlap between objects stops non-max suppression from working
Y LeCunImageNet 2013: Detection
200 categories. System must give 5 bounding boxes with categories
OverFeat: pre-trained on ImageNet 1K, fine-tuned on ImageNet Detection.
Post-deadline result: 0.243 mean average precision (best known result until May 2014)
Y LeCun
Results: pre-trained on ImageNet1K,fine-tuned on ImageNet Detection
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection Examples
Y LeCunDetection: Difficult Examples
Groundtruth is sometimes ambiguous or incomplete
Y LeCunDetection: Difficult Examples
Non-max suppression makes us miss many objects
Person behind instrument
A bit of contextual post-processing would fix many errors
Y LeCunDetection: Interesting Failures
Snake → Corkscrew
Y LeCunDetection: Bad Groundtruth
One of the labelers likes ticks.....
Y LeCun
State of the art with only 6 training examples
Features are generic: Caltech 256
Network first trained on ImageNet.
Last layer chopped off
Last layer trained on Caltech 256,
first layers N-1 kept fixed.
State of the art accuracy with only 6 training samples/class
3: [Bo, Ren, Fox. CVPR, 2013] 16: [Sohn, Jung, Lee, Hero ICCV 2011]
Y LeCunCats vs Dogs
● Kaggle competition: Dog vs Cats
Y LeCunCats vs Dogs
● Won by Pierre Sermanet (NYU):● ImageNet network (OverFeat) last layers retrained on cats and dogs
Y LeCunReal-Time Demo
Real-Time demo2.6 fps on quadcore Intel7.6fps on Nvidia Geforce GTX 680m (mobile GPU)
Network trained in ImageNet 1K
Y LeCun
Other ApplicationsOf
Convolutional Nets
Y LeCun
ConvNet in Connectomics [Jain, Turaga, Seung 2007-present]
3D ConvNet
Volumetric
Images
Each voxel labeled as “membrane” or “non-membrane using a 7x7x7 voxel neighborhood
Y LeCunFace Recognition:DeepFace (Facebook AI Research)
[Taigman et al. CVPR 2014]AlignmentConvnet
Y LeCunFace Recognition:DeepFace (Facebook AI Research)
Performanceon Labeled Face in the Wild dataset (LFW)
Y LeCunDeepFace: performance
Y LeCunPose Estimation and Attribute Recovery with ConvNets
Body pose estimation [Tompson et al. ICLR, 2014]
Real-time hand pose recovery
[Tompson et al. Trans. on Graphics 14]
Pose-Aligned Network for Deep Attribute Modeling
[Zhang et al. CVPR 2014]
Y LeCunOverFeat Features ->Trained Classifier on other datasets
A. S. Razavian , H. Azizpour , J. Sullivan , S. Carlsson "CNN features off-the-shelf: An astounding baseline for recogniton", CVPR 2014, DeepVision Workshop.
http://www.csc.kth.se/cvap/cvg/DL/ots/
Y LeCunOverFeat Features + Classifer on various datasets
Y LeCunOther ConvNet Results
Y LeCunOther ConvNet Results
Results compiled by Pierre Sermanethttp://cs.nyu.edu/~sermanet/papers/Deep_ConvNets_for_Vision-Results.pdf
Y LeCunOther Tasks for Which Deep Convolutional Nets are the Best
Handwriting recognition MNIST (many), Arabic HWX (IDSIA)OCR in the Wild [2011]: StreetView House Numbers (NYU and others)Traffic sign recognition [2011] GTSRB competition (IDSIA, NYU)Asian handwriting recognition [2013] ICDAR competition (IDSIA)Pedestrian Detection [2013]: INRIA datasets and others (NYU)Volumetric brain image segmentation [2009] connectomics (IDSIA, MIT)Human Action Recognition [2011] Hollywood II dataset (Stanford)Object Recognition [2012] ImageNet competition (Toronto)Scene Parsing [2012] Stanford bgd, SiftFlow, Barcelona datasets (NYU) Scene parsing from depth images [2013] NYU RGB-D dataset (NYU)Speech Recognition [2012] Acoustic modeling (IBM and Google)Breast cancer cell mitosis detection [2011] MITOS (IDSIA)
The list of perceptual tasks for which ConvNets hold the record is growing.Most of these tasks (but not all) use purely supervised convnets.
Y LeCun
Deep Learning andConvolutional Networks inSpeech, Audio, and Signals
Y LeCun
Feature
Extraction
NeuralNetwork
Decoder
Transducer&
LanguageModel
Hi, how are you?Acoustic Modeling in Speech Recognition (Google)
A typical speech recognition architecture with DL-based acoustic modelingFeatures: log energy of a filter bank (e.g. 40 filters)Neural net acoustic modeling (convolutional or not)Input window: typically 10 to 40 acoustic framesFully-connected neural net: 10 layers, 2000-4000 hidden units/layerBut convolutional nets do better....Predicts phone state, typically 2000 to 8000 categories
Mohamed et al. “DBNs for phone recognition” NIPS Workshop 2009Zeiler et al. “On rectified linear units for speech recognition” ICASSP 2013
Y LeCunSpeech Recognition with Convolutional Nets (NYU/IBM)
Acoustic Model: ConvNet with 7 layers. 54.4 million parameters.
Classifies acoustic signal into 3000 context-dependent subphones categories
ReLU units + dropout for last layers
Trained on GPU. 4 days of training
Y LeCunSpeech Recognition with Convolutional Nets (NYU/IBM)
Subphone-level classification error (sept 2013):Cantonese: phone: 20.4% error; subphone: 33.6% error (IBM DNN: 37.8%)
Subphone-level classification error (march 2013)Cantonese: subphone: 36.91%Vietnamese: subphone 48.54%Full system performance (token error rate on conversational speech):76.2% (52.9% substitution, 13.0% deletion, 10.2% insertion)
Y LeCunSpeech Recognition with Convolutional Nets (NYU/IBM)
Training samples. 40 MEL-frequency Cepstral CoefficientsWindow: 40 frames, 10ms each
Y LeCunSpeech Recognition with Convolutional Nets (NYU/IBM)
Convolution Kernels at Layer 1:64 kernels of size 9x9
Y LeCun
Convolutional NetworksIn
Semantic Segmentation,Scene Labeling
Y LeCun
Semantic Labeling:Labeling every pixel with the object it belongs to
[Farabet et al. ICML 2012, PAMI 2013]
Would help identify obstacles, targets, landing sites, dangerous areasWould help line up depth map with edge maps
Y LeCunScene Parsing/Labeling: ConvNet Architecture
Each output sees a large input context:46x46 window at full rez; 92x92 at ½ rez; 184x184 at ¼ rez
[7x7conv]->[2x2pool]->[7x7conv]->[2x2pool]->[7x7conv]->
Trained supervised on fully-labeled images
Laplacian
Pyramid
Level 1
Features
Level 2
Features
Upsampled
Level 2 Features
Categories
Y LeCun
Method 1: majority over super-pixel regions
[Farabet et al. IEEE T. PAMI 2013]M
ulti-sca le ConvN
etSuper-pix el bound ary hype theses
Convolut ional clas sifier
Majority
Vote
Over
Superpixels
Input image
Superpixel boundaries
Features from
Convolutional net
(d=768 per pixel)
“soft” categories scores
Categories aligned
With region
boundaries
Y LeCunScene Parsing/Labeling: Performance
Stanford Background Dataset [Gould 1009]: 8 categories
[Farabet et al. IEEE T. PAMI 2013]
Y LeCunScene Parsing/Labeling: Performance
[Farabet et al. IEEE T. PAMI 2012]
SIFT Flow Dataset[Liu 2009]: 33 categories
Barcelona dataset[Tighe 2010]: 170 categories.
Y LeCunScene Parsing/Labeling: SIFT Flow dataset (33 categories)
Samples from the SIFT-Flow dataset (Liu)
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling: SIFT Flow dataset (33 categories)
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling
[Farabet et al. ICML 2012, PAMI 2013]
Y LeCunScene Parsing/Labeling
No post-processingFrame-by-frameConvNet runs at 50ms/frame on Virtex-6 FPGA hardware
But communicating the features over ethernet limits system performance
Y LeCunTemporal Consistency
Spatio-Temporal Super-Pixel segmentation [Couprie et al ICIP 2013][Couprie et al JMLR under review]Majority vote over super-pixels
Y LeCunScene Parsing/Labeling: Temporal Consistency
Causal method for temporal consistency
[Couprie, Farabet, Najman, LeCun ICLR 2013, ICIP 2013]
Y LeCunNYU RGB-D Dataset
Captured with a Kinect on a steadycam
Y LeCunResults
Depth helps a bitHelps a lot for floor and propsHelps surprisingly little for structures, and hurts for furniture
[C. Cadena, J. Kosecka “Semantic Parsing for Priming Object Detection in RGB-D Scenes”Semantic Perception Mapping and Exploration (SPME), Karlsruhe 2013]
Y LeCunScene Parsing/Labeling on RGB+Depth Images
With temporal consistency
[Couprie, Farabet, Najman, LeCun ICLR 2013, ICIP 2013]
Y LeCunScene Parsing/Labeling on RGB+Depth Images
With temporal consistency
[Couprie, Farabet, Najman, LeCun ICLR 2013, ICIP 2013]
Y LeCunLabeling Videos
Temporal consistency
[Couprie, Farabet, Najman, LeCun ICLR 2013][Couprie, Farabet, Najman, LeCun ICIP 2013][Couprie, Farabet, Najman, LeCun submitted to JMLR]
Y LeCunSemantic Segmentation on RGB+D Images and Videos
[Couprie, Farabet, Najman, LeCun ICLR 2013, ICIP 2013]
Y LeCunCommercial Applications of Convolutional Nets
Form Reading: AT&T 1994
Check reading: AT&T 1996 (read 10-20% of all US checks in 2000)
Handwriting recognition: Microsoft early 2000
Face and person detection: NEC 2005
Gender and age recognition: NEC 2010 (vending machines)
OCR in natural images: Google 2013 (StreetView house numbers)
Photo tagging: Google 2013
Image Search by Similarity: Baidu 2013
Many applications at Facebook, Google, Baidu, Microsoft, IBM, NEC,..... Speech recognition, face recognition, image search, porn filtering,....
Y LeCun
Self-LearningConvNet for Scene Labeling
Vision-Based Navigation
for Off-Road Robots
Y LeCunLAGR project: vision-based navigation for off-road robot
Getting a robot to drive autonomously in unknown terrain solely from vision (camera input).
Our team (NYU/Net-Scale Technologies Inc.) was one of 8 participants funded by DARPA
All teams received identical robots and can only modify the software (not the hardware)
The robot is given the GPS coordinates of a goal, and must drive to the goal as fast as possible. The terrain is unknown in advance. The robot is run 3 times through the same course.
Long-Range Obstacle Detection with on-line, self-trained ConvNet
Uses temporal consistency!
Y LeCunObstacle Detection at Short Range: Stereovision
Camera image Detected obstacles (red)
Obstacles overlaid with camera image
Y LeCunBut Stereovision Doesn't work at long range
Stereo is only good up to about 10 meters.
But not seeing past 10 meters is like driving in a fog or a snowstorm!
Y LeCunLong Range Vision with a Convolutional Net
Pre-processing (125 ms)– Ground plane estimation– Horizon leveling– Conversion to YUV + local
contrast normalization– Scale invariant pyramid of
distance-normalized image “bands”
Y LeCun
Input imageInput image Stereo LabelsStereo Labels
Classifier OutputClassifier Output
Scene Labeling with ConvNet + online learning
Image Labeling for Off-Road Robots [Hadsell JFR 2008]ConvNet labels pixels as one of 3 categoriesTraversible/flat (green), non traversible (red), foot of obstacle (purple)Labels obtained from stereo vision and SLAM
Input imageInput image Stereo LabelsStereo Labels Classifier OutputClassifier Output
Y LeCunLong Range Vision Results
Input imageInput image Stereo LabelsStereo Labels Classifier OutputClassifier Output
Input imageInput image Stereo LabelsStereo Labels Classifier OutputClassifier Output
Y LeCunLong Range Vision Results
Input imageInput image Stereo LabelsStereo Labels Classifier OutputClassifier Output
Input imageInput image Stereo LabelsStereo Labels Classifier OutputClassifier Output
Y LeCun
Software Tools and
Hardware Accelerationfor
Convolutional Networks
Y LeCunModule-Based Deep Learning with Torch7
Torch7 is based on the Lua languageSimple and lightweight scripting language, dominant in the game industryHas a native just-in-time compiler (fast!)Has a simple foreign function interface to call C/C++ functions from Lua
Torch7 is an extension of Lua withA multidimensional array engine with CUDA and OpenMP backendsA machine learning library that implements multilayer nets, convolutional nets, unsupervised pre-training, etcVarious libraries for data/image manipulation and computer visionA quickly growing community of users
Single-line installation on Ubuntu and Mac OSX:http://torch.ch
Torch7 Cheat sheet (with links to libraries and tutorials):– https://github.com/torch/torch7/wiki/Cheatsheet
Y LeCunExample: building a Neural Net in Torch7
Net for SVHN digit recognition
10 categories
Input is 32x32 RGB (3 channels)
1500 hidden units
Creating a 2-layer net
Make a cascade module
Reshape input to vector
Add Linear module
Add tanh module
Add Linear Module
Add log softmax layer
Create loss function module
Noutputs = 10; nfeats = 3; Width = 32; height = 32ninputs = nfeats*width*heightnhiddens = 1500
Simple 2layer neural networkmodel = nn.Sequential()model:add(nn.Reshape(ninputs))model:add(nn.Linear(ninputs,nhiddens))model:add(nn.Tanh())model:add(nn.Linear(nhiddens,noutputs))model:add(nn.LogSoftMax())
criterion = nn.ClassNLLCriterion()
See Torch7 example at http://bit.ly/16tyLAx
Y LeCunSpecialized Hardware can Accelerate Convolutions
Large-scale convolutional networks are trained on GPUsGenerally implemented on NVIDIA GPUs using CUDABut exploiting all the power of GPUs for ConvNets is extremely difficultLimited by memory bandwidth and memory architecture
Multi-core architectures may become competitivee.g. Intel Xeon Phi
Major hardware manufacturers are exploring how to support convolutional nets more efficiently with their hardware.
Direct support for convnet operations in mobile CPU/GPU, as well as in high-end CPU/GPU may soon become available.
But dedicated ConvNet hardware is also on the wayMostly for embedded applications (smart cameras, robots...)
Y LeCunNeuFlow architecture (NYU + Purdue)
Collaboration NYU-Purdue: Eugenio Culurciello's e-Lab.
Running on Picocomputing 8x10cm high-performance FPGA boardVirtex 6 LX240T: 680 MAC units, 20 neuflow tiles
Full scene labeling at 20 frames/sec (50ms/frame) at 320x240
board with Virtex-6
Y LeCunNewFlow: Architecture
grid of passive processing tiles (PTs)
CPU
DMA
MEM
global networkonchip to allow fast reconfiguration
RISC CPU, to reconfigure tiles and data paths, at runtime
Multiport memory controller (DMA)
[x20 on a Virtex6 LX240T]
[x12 on a V6 LX240T]
Y LeCunNewFlow: Processing Tile Architecture
Termby:term streaming operators (MUL,DIV,ADD,SUB,MAX)
configurable bank of FIFOs , for stream buffering, up to 10kB per PT
full 1/2D parallel convolver with 100 MAC units
configurable piecewise linear or quadratic mapper
configurable router, to stream data in and out of the tile, to neighbors or DMA ports
[x8,2 per tile]
[x4] [x8]
[x4]
[Virtex6 LX240T]
[x20]
Y LeCunNewFlow ASIC: 2.5x5 mm, 45nm, 0.6Watts, >300GOPS
Collaboration Purdue-NYU: Eugenio Culurciello's e-Lab
Suitable for vision-enabled embedded and mobile devices
(but the fabrication was botched...)
[Pham, Jelaca, Farabet, Martini, LeCun, Culurciello 2012]
Y LeCunNewFlow: Performance
IntelIntelI7 4 coresI7 4 cores
neuFlowneuFlowVirtex4Virtex4
neuFlowneuFlowVirtex 6Virtex 6
nVidia nVidia GT335mGT335m
NeuFlowNeuFlowASIC ASIC 45nm45nm
nVidianVidiaGTX480*GTX480*
PeakGOP/sec 40 40 160 182 320 1350Actual
GOP/sec 12 37 147 54 300 294
FPS 14 46 182 67 364 374Power (W) 50 10 10 30 0.6 220
Embed?(GOP/s/W) 0.24 3.7 14.7 1.8 490 1.34
NeuFlow Virtex6 can run the semantic labeling system at 50ms/frame* performance of Nvidia GPU is higher when using minibatch training
Y LeCun
Unsupervised Learning
Y LeCun Sparse auto-encoder: Predictive Sparse Decomposition (PSD)
Prediction the optimal code with a trained encoder
Energy = reconstruction_error + code_prediction_error + code_sparsity
E Y i , Z =∥Y i−W d Z∥
2∥Z−ge W e ,Y i
∥2∑ j
∣z j∣
ge (W e , Y i)=shrinkage(W e Y i
)
[Kavukcuoglu, Ranzato, LeCun, 2008 → arXiv:1010.3467],
INPUT Y Z
∥Y i− Y∥
2
∣z j∣
W d Z
FEATURES
∑ j.
∥Z− Z∥2ge W e ,Y i
Y LeCun
PSD: Basis Functions on MNIST
Basis functions (and encoder matrix) are digit parts
Y LeCun
Training on natural images patches.
12X12256 basis functions
Predictive Sparse Decomposition (PSD): Training
Y LeCun
Learned Features on natural patches: V1-like receptive fields
Y LeCun
Replace the dot products with dictionary element by convolutions.Input Y is a full imageEach code component Zk is a feature map (an image)Each dictionary element is a convolution kernel
Regular sparse coding
Convolutional S.C.
∑k
. * Zk
Wk
Y =
“deconvolutional networks” [Zeiler, Taylor, Fergus CVPR 2010]
Convolutional Sparse Coding
Y LeCun
Convolutional FormulationExtend sparse coding from PATCH to IMAGE
PATCH based learning CONVOLUTIONAL learning
Convolutional PSD: Encoder with a soft sh() Function
Y LeCunConvolutional Sparse Auto-Encoder on Natural Images
Filters and Basis Functions obtained with 1, 2, 4, 8, 16, 32, and 64 filters.
Y LeCun
Phase 1: train first layer using PSD
FEATURES
Y Z
∥Y i−Y∥2
∣z j∣
W d Z λ∑ .
∥Z−Z∥2g e (W e ,Yi)
Using PSD to Train a Hierarchy of Features
Y LeCun
Phase 1: train first layer using PSD
Phase 2: use encoder + absolute value as feature extractor
FEATURES
Y ∣z j∣
g e (W e ,Yi)
Using PSD to Train a Hierarchy of Features
Y LeCun
Phase 1: train first layer using PSD
Phase 2: use encoder + absolute value as feature extractor
Phase 3: train the second layer using PSD
FEATURES
Y ∣z j∣
g e (W e ,Yi)
Y Z
∥Y i−Y∥2
∣z j∣
W d Z λ∑ .
∥Z−Z∥2g e (W e ,Yi)
Using PSD to Train a Hierarchy of Features
Y LeCun
Phase 1: train first layer using PSD
Phase 2: use encoder + absolute value as feature extractor
Phase 3: train the second layer using PSD
Phase 4: use encoder + absolute value as 2nd feature extractor
FEATURES
Y ∣z j∣
g e (W e ,Yi)
∣z j∣
g e (W e ,Yi )
Using PSD to Train a Hierarchy of Features
Y LeCun
Phase 1: train first layer using PSD
Phase 2: use encoder + absolute value as feature extractor
Phase 3: train the second layer using PSD
Phase 4: use encoder + absolute value as 2nd feature extractor
Phase 5: train a supervised classifier on top
Phase 6 (optional): train the entire system with supervised back-propagation
FEATURES
Y ∣z j∣
g e (W e ,Yi)
∣z j∣
g e (W e ,Yi )
classifier
Using PSD to Train a Hierarchy of Features
Y LeCun
Unsupervised + SupervisedFor
Pedestrian Detection
Y LeCun
[Osadchy,Miller LeCun JMLR 2007],[Kavukcuoglu et al. NIPS 2010] [Sermanet et al. CVPR 2013]
Pedestrian Detection, Face Detection
Y LeCun
Feature maps from all stages are pooled/subsampled and sent to the final classification layers
Pooled low-level features: good for textures and local motifsHigh-level features: good for “gestalt” and global shape
[Sermanet, Chintala, LeCun CVPR 2013]
7x7 filter+tanh
38 feat maps
Input
78x126xYUV
L2 Pooling
3x3
2040 9x9
filters+tanh
68 feat maps
Av Pooling
2x2 filter+tanh
ConvNet Architecture with Multi-Stage Features
Y LeCun
[Kavukcuoglu et al. NIPS 2010] [Sermanet et al. ArXiv 2012]
ConvNet
Color+Skip
Supervised
ConvNet
Color+Skip
Unsup+Sup
ConvNet
B&W
Unsup+Sup
ConvNet
B&W
Supervised
Pedestrian Detection: INRIA Dataset. Miss rate vs false positives
Y LeCun
128 stage-1 filters on Y channel.
Unsupervised training with convolutional predictive sparse decomposition
Unsupervised pre-training with convolutional PSD
Y LeCun
Stage 2 filters.
Unsupervised training with convolutional predictive sparse decomposition
Unsupervised pre-training with convolutional PSD
Y LeCun
VIDEOS
Y LeCun
VIDEOS
Y LeCun
[Kavukcuoglu et al. NIPS 2010] [Sermanet et al. ArXiv 2012]
ConvNet
Color+Skip
Supervised
ConvNet
Color+Skip
Unsup+Sup
ConvNet
B&W
Unsup+Sup
ConvNet
B&W
Supervised
Pedestrian Detection: INRIA Dataset. Miss rate vs false positives
Y LeCun
Unsupervised Learning:Invariant Features
Y LeCunLearning Invariant Features with L2 Group Sparsity
Unsupervised PSD ignores the spatial pooling step.Could we devise a similar method that learns the pooling layer as well?Idea [Hyvarinen & Hoyer 2001]: group sparsity on pools of features
Minimum number of pools must be non-zero
Number of features that are on within a pool doesn't matter
Pools tend to regroup similar features
INPUT Y Z
∥Y i−Y∥2 W d Z
FEATURES
λ∑ .
∥Z−Z∥2g e (W e ,Yi )
√ (∑ Z k2 )
L2 norm within each pool
E (Y,Z )=∥Y−W d Z∥2+∥Z−g e (W e ,Y )∥
2+∑j √ ∑
k∈P j
Z k2
Y LeCun
Learning Invariant Features with L2 Group Sparsity
Idea: features are pooled in group. Sparsity: sum over groups of L2 norm of activity in group.
[Hyvärinen Hoyer 2001]: “subspace ICA” decoder only, square
[Welling, Hinton, Osindero NIPS 2002]: pooled product of experts encoder only, overcomplete, log student-T penalty on L2 pooling
[Kavukcuoglu, Ranzato, Fergus LeCun, CVPR 2010]: Invariant PSDencoder-decoder (like PSD), overcomplete, L2 pooling
[Le et al. NIPS 2011]: Reconstruction ICASame as [Kavukcuoglu 2010] with linear encoder and tied decoder
[Gregor & LeCun arXiv:1006:0448, 2010] [Le et al. ICML 2012]Locally-connect non shared (tiled) encoder-decoder
INPUT
YEncoder only (PoE, ICA),
Decoder Only or
Encoder-Decoder (iPSD, RICA)Z INVARIANT
FEATURES
λ∑ .
√ (∑ Z k2 )
L2 norm within each pool
SIMPLE FEATURES
Y LeCunGroups are local in a 2D Topographic Map
The filters arrange themselves spontaneously so that similar filters enter the same pool.The pooling units can be seen as complex cellsOutputs of pooling units are invariant to local transformations of the input
For some it's translations, for others rotations, or other transformations.
Y LeCunImage-level training, local filters but no weight sharing
Training on 115x115 images. Kernels are 15x15 (not shared across space!)
Y LeCun
119x119 Image Input100x100 Code
20x20 Receptive field sizesigma=5 Michael C. Crair, et. al. The Journal of Neurophysiology
Vol. 77 No. 6 June 1997, pp. 3381-3385 (Cat)
K Obermayer and GG Blasdel, Journal of Neuroscience, Vol 13, 4114-4129 (Monkey)Topographic Maps
Y LeCunImage-level training, local filters but no weight sharing
Color indicates orientation (by fitting Gabors)
Y LeCunInvariant Features Lateral Inhibition
Replace the L1 sparsity term by a lateral inhibition matrixEasy way to impose some structure on the sparsity
[Gregor, Szlam, LeCun NIPS 2011]
Y LeCunInvariant Features via Lateral Inhibition: Structured Sparsity
Each edge in the tree indicates a zero in the S matrix (no mutual inhibition)
Sij is larger if two neurons are far away in the tree
Y LeCunInvariant Features via Lateral Inhibition: Topographic Maps
Non-zero values in S form a ring in a 2D topologyInput patches are high-pass filtered
Y LeCunInvariant Features through Temporal Constancy
Object is cross-product of object type and instantiation parametersMapping units [Hinton 1981], capsules [Hinton 2011]
small medium large
Object type Object size[Karol Gregor et al.]
Y LeCunWhat-Where Auto-Encoder Architecture
St St-1 St-2
C1t C
1t-1 C
1t-2 C
2t
Decoder
W1 W1 W1 W2
Predictedinput
C1t C
1t-1 C
1t-2 C
2t
St St-1 St-2
Inferred code
Predictedcode
InputEncoder
f ∘ W 1 f ∘ W 1 f ∘ W 1
W 2
f
W 2
W 2
Y LeCunLow-Level Filters Connected to Each Complex Cell
C1(where)
C2(what)
Y LeCun
Input
Generating Images
Generating images
Y LeCun
Deep LearningAnd
Structured Prediction
Y LeCunIntegrating Deep Learning and Structured Prediction
Structured prediction: when the output is structured: string, graph.....
Integrating deep learning and structured prediction is an old idea
In fact, it predates structured prediction [LeCun, Bottou, Bengio, Haffner 1998]Globally-trained convolutional-net + graphical models for handwriting recognition
trained discriminatively at the word level
Loss identical to CRF and structured perceptron
Compositional movable parts model
A system like this was reading 10 to 20% of all the checks in the US around 1998
Y LeCunDeep Recurrent Neural Language Model
[Mirowski-LeCun MLKDD 2009][Mirowski et al.
Y LeCun
FutureChallenges
Y LeCunFuture Challenges
Integrated feed-forward and feedbackDeep Boltzmann machine do this, but there are issues of scalability.
Integrating supervised and unsupervised learning in a single algorithmAgain, deep Boltzmann machines do this, but....
Integrating deep learning and structured prediction (“reasoning”)This has been around since the 1990's but needs to be revived
Learning representations for complex reasoning“recursive” networks that operate on vector space representations of knowledge [Pollack 90's] [Bottou 2010] [Socher, Manning, Ng 2011]
Representation learning in natural language processing[Y. Bengio 01],[Collobert Weston 10], [Mnih Hinton 11] [Socher 12]
Better theoretical understanding of deep learning and convolutional netse.g. Stephane Mallat's “scattering transform”, work on the sparse representations from the applied math community....
Y LeCunTowards Practical AI: Challenges
Applying deep learning to NLP (requires “structured prediction”)
Video analysis/understanding (requires unsupervised learning)
High-performance/low power embedded systems for ConvNets (FPGA/ASIC?)
Very-large-scale deep learning (distributed optimization)
Integrating reasoning with DL (“energy-based models”, recursive neural nets)
Then we can haveAutomatically-created high-performance data analytics systemsVector-space embedding of everything (language, users,...)Multimedia content understanding, search and indexingMultilingual speech dialog systemsDriver-less carsAutonomous maintenance robots / personal care robots
Y LeCunThe Graph of Deep Learning Sparse Modeling Neuroscience↔ ↔
Architecture of V1
[Hubel, Wiesel 62]
Basis/Matching Pursuit
[Mallat 93; Donoho 94]
Sparse Modeling
[Olshausen-Field 97]
Neocognitron
[Fukushima 82]Backprop
[many 85]
Convolutional Net
[LeCun 89]
Sparse Auto-Encoder
[LeCun 06; Ng 07]
Restricted
Boltzmann
Machine
[Hinton 05]
Normalization
[Simoncelli 94]
Speech Recognition
[Goog, IBM, MSFT 12]
Object Recog
[Hinton 12]Scene Labeling
[LeCun 12]
Connectomics
[Seung 10]
Object Reco
[LeCun 10]
Compr. Sensing
[Candès-Tao 04]
L2-L1 optim
[Nesterov,
Nemirovski
Daubechies,
Osher....]
Scattering
Transform
[Mallat 10]
Stochastic Optimization
[Nesterov, Bottou
Nemirovski,....]
Sparse Modeling
[Bach, Sapiro. Elad]MCMC, HMC
Cont. Div.
[Neal, Hinton]
Visual Metamers
[Simoncelli 12]